AMD CPUs, SoC Rumors and Speculations Temp. thread 2

[juanrga]

The SPARC M7 ships at 4.1Ghz with 32 cores. There is little reason the 14/16nm parts which use a similar 20nm BEOL can't hit similar targets with just 8 cores.

TSMC 20HP != Samsung/Glofo 14LPE/LPP

HP means High Performance

LPE means Low Power Early

LPP means Low Power Plus

 

[juanrga]

The 28SHP process used in Kaveri cannot hit frequencies so higher as 32PDSOI because is a process node optimized for density not performance.

The 14LPP process node cannot hit frequencies so higher as 32PDSOI because is a process node optimized for mobile applications. Phones run at sub 3GHz; therefore, it is likely to think that 14LPP is not optimized for higher frequencies.

We could maybe expect ~2.8GHz base and ~3.5GHz turbo. What is beyond any reasonable doubt is the possibility that Zen hits 4GHz base.

I asked him about turbos. His reply was "8000-series Carrizo like." In the past he provided us turbo frequencies for Carrizo before release, and he was very close. I am going to trust him on Zen as well.

 

[salgado18]

The 28SHP process used in Kaveri cannot hit frequencies so higher as 32PDSOI because is a process node optimized for density not performance.

The 14LPP process node cannot hit frequencies so higher as 32PDSOI because is a process node optimized for mobile applications. Phones run at sub 3GHz; therefore, it is likely to think that 14LPP is not optimized for higher frequencies.

We could maybe expect ~2.8GHz base and ~3.5GHz turbo. What is beyond any reasonable doubt is the possibility that Zen hits 4GHz base.

I asked him about turbos. His reply was "8000-series Carrizo like." In the past he provided us turbo frequencies for Carrizo before release, and he was very close. I am going to trust him on Zen as well.

For those of us (like me) who don't know clocks of newer cpus, the Carrizo line has three of them:

FX-8800P: 2.1GHz base, 3.4GHz turbo
A10-8700P: 1.8GHz base, 3.2GHz turbo
A8-8600P: 1.6GHz base, 3.0GHz turbo

I think that's too low for a desktop CPU, especially with Ivy-Haswell performance levels. Could AMD be trying the moar-cores game again? Or could they try to stretch the TDP to reach higher clocks?

EDIT: Wait, I just remember a slide from AMD about Carrizo, stating there would be gains over Kaveri in both IPC and clocks. But the top Kaveri has clock speeds of 2.7/3.6 GHz. Could they be still developing better Carrizos with higher clocks? If so, does he mean current Carrizo, or in-development Carrizo?

 

[gamerk316]

I'd imagine the base would be higher, but a turbo up to 3.4 GHz or so wouldn't shock me.

 

[Vogner16]

http://www.tomshardware.com/forum/id-2864699/zen-theoritical-performance.html

my testing

 

[Vogner16]

BD was a speed-demon microarchitecture. The goal was to hit higher frequencies at cost of IPC. BD did initially target too high frequencies and the Glofo process node couldn't support them. Only recently a mature 32PDSOI and certain microarchitecture tweaks allowed Piledriver to hit 5GHz turbo. That is still far from the maximum theoretical allowed by the microarchitecture: ~10GHz.

The 28SHP process used in Kaveri cannot hit frequencies so higher as 32PDSOI because is a process node optimized for density not performance.

The 14LPP process node cannot hit frequencies so higher as 32PDSOI because is a process node optimized for mobile applications. Phones run at sub 3GHz; therefore, it is likely to think that 14LPP is not optimized for higher frequencies.

We could maybe expect ~2.8GHz base and ~3.5GHz turbo. What is beyond any reasonable doubt is the possibility that Zen hits 4GHz base.

I highly doubt that zen would use a standard Glofo 14LPP process. they have always used specialized processes in the past as you know the 28nm process amd uses for Carrizo is different than the process used by early 28nm chips labled as low power /mobile application. expect something like 14HPP as LPP is designed for sub 35W chips and I expect a roughly 100W tdp for 14nm parts, possibly 150W for the 16 thread monster.

 

[8350rocks]

The SPARC M7 ships at 4.1Ghz with 32 cores. There is little reason the 14/16nm parts which use a similar 20nm BEOL can't hit similar targets with just 8 cores.

TSMC 20nm high performance node != Samsung/Glofo 14LPE/LPP

You do realize that the process AMD is using is a tuned version of the samsung node...right?

This is not just off the shelf 14nm LPP. I also question what "insider foundry knowledge" others would claim to have...as 70% sounds quite a bit off of what I am hearing...but whatever.

 

[jimmysmitty]

BD was a speed-demon microarchitecture. The goal was to hit higher frequencies at cost of IPC. BD did initially target too high frequencies and the Glofo process node couldn't support them. Only recently a mature 32PDSOI and certain microarchitecture tweaks allowed Piledriver to hit 5GHz turbo. That is still far from the maximum theoretical allowed by the microarchitecture: ~10GHz.

The 28SHP process used in Kaveri cannot hit frequencies so higher as 32PDSOI because is a process node optimized for density not performance.

The 14LPP process node cannot hit frequencies so higher as 32PDSOI because is a process node optimized for mobile applications. Phones run at sub 3GHz; therefore, it is likely to think that 14LPP is not optimized for higher frequencies.

We could maybe expect ~2.8GHz base and ~3.5GHz turbo. What is beyond any reasonable doubt is the possibility that Zen hits 4GHz base.

I highly doubt that zen would use a standard Glofo 14LPP process. they have always used specialized processes in the past as you know the 28nm process amd uses for Carrizo is different than the process used by early 28nm chips labled as low power /mobile application. expect something like 14HPP as LPP is designed for sub 35W chips and I expect a roughly 100W tdp for 14nm parts, possibly 150W for the 16 thread monster.

Even with a modified node Zen will be dropping clock speed much like Core 2 did vs Pentium 4. Before Core 2 they were at 3.8GHz-4GHz top end speeds but Core 2 was topping out at 3GHz on the same node. 45nm got them to over 3GHz stock and 4GHz overclocks though

That is due to the change in the pipelines more than anything. NetBurst was a long pipeline core much like Bulldozer so it could reach pretty high clock speeds (it was supposed to be able to hit 10GHz also). Zen will probably have shorter but more efficient pipelines which is a good thing. Clock speed is not the end all be all, AMD proved that with Athlon 64 and again Intel proved it with Core 2. Hell most i5/i7s at stock speed are more than enough to push most applications. Any overclocking is now for epeen points.

 

[Vogner16]

I remember overclocking my core 2 duo to 3.5 😛 ahh windows vista how we don't miss you

aside from that your right. clock speed isn't everything, but for zen to match overall performance of intel they need to match or exceed clock speeds as their IPC will not be as good.

 

[etayorius]

Should we already have an Exact release date by now? something tells me it may not even be available till 2017.


I don`t care the clock speed, as long as it at least competes with Haswell... for all i care it could be clocked to 2GHZ max.

 

[jimmysmitty]

Should we already have an Exact release date by now? something tells me it may not even be available till 2017.

I would think we should have at least a quarter release date, not the exact date, by now. But then again AMD has been pretty closed lipped for most things.

 

[cdrkf]

Thinking about the 'carrizo like turbo' I would suggest that refers to range rather than specific speeds (as carrizo is a much lower power part).

I mean carrizo boosts by over 1ghz. That would fit with Zen having a low base when all cores loaded (e.g. like suggested 2.8ghz) but a huge boost range pushing it up to the 4ghz range on lighter workloads.

That would actually work well, as older low thread count games would allow it to push a couple of cores way up giving good frame rates, whilst the high core count provides benefit in threaded workloads even at lower speeds.

 

[salgado18]

Thinking about the 'carrizo like turbo' I would suggest that refers to range rather than specific speeds (as carrizo is a much lower power part).

I mean carrizo boosts by over 1ghz. That would fit with Zen having a low base when all cores loaded (e.g. like suggested 2.8ghz) but a huge boost range pushing it up to the 4ghz range on lighter workloads.

That would actually work well, as older low thread count games would allow it to push a couple of cores way up giving good frame rates, whilst the high core count provides benefit in threaded workloads even at lower speeds.

Also, we can't forget that 8-core Intels also have low base and turbo clocks. The ones at 4.0 range are the 4-core processors. If the source says "at 8000-Carrizo level", does he mean the 8-core, the fast ones like the 4-cores, or the entire range will have around that?

 

[juanrga]

The 28SHP process used in Kaveri cannot hit frequencies so higher as 32PDSOI because is a process node optimized for density not performance.

The 14LPP process node cannot hit frequencies so higher as 32PDSOI because is a process node optimized for mobile applications. Phones run at sub 3GHz; therefore, it is likely to think that 14LPP is not optimized for higher frequencies.

We could maybe expect ~2.8GHz base and ~3.5GHz turbo. What is beyond any reasonable doubt is the possibility that Zen hits 4GHz base.

I asked him about turbos. His reply was "8000-series Carrizo like." In the past he provided us turbo frequencies for Carrizo before release, and he was very close. I am going to trust him on Zen as well.

For those of us (like me) who don't know clocks of newer cpus, the Carrizo line has three of them:

FX-8800P: 2.1GHz base, 3.4GHz turbo
A10-8700P: 1.8GHz base, 3.2GHz turbo
A8-8600P: 1.6GHz base, 3.0GHz turbo

I think that's too low for a desktop CPU, especially with Ivy-Haswell performance levels. Could AMD be trying the moar-cores game again? Or could they try to stretch the TDP to reach higher clocks?

EDIT: Wait, I just remember a slide from AMD about Carrizo, stating there would be gains over Kaveri in both IPC and clocks. But the top Kaveri has clock speeds of 2.7/3.6 GHz. Could they be still developing better Carrizos with higher clocks? If so, does he mean current Carrizo, or in-development Carrizo?

AMD Zen FX CPU tops out at 8-core; Broadwell-E CPU tops out at 10-core. The only place where AMD seems to push a moar cores strategy is on servers: Zen Opteron up to 32-core; Broadwell Xeon up to 24-core.

AMD Zen FX CPU is rated up to 95W. Haswell-E is up to 140W. Broadwell-E seems to maintain same wattage.

AMD is not developing new Carrizos. Only the lower rated Carrizos achieve higher clocks than Kaveri parts. The top Carrizos achieve lower clocks than Kaveri, because Carrizo uses HDL which reduces maximum clocks.

The crossing point is about 20W. Higher wattage Carrizos are less efficient and achieve lower clocks compared to Kaveri. Lower wattage Carrizos are more efficient and achieve higher clocks than Kaveri parts.

 

[juanrga]

You do realize that the process AMD is using is a tuned version of the samsung node...right?

This is not just off the shelf 14nm LPP. I also question what "insider foundry knowledge" others would claim to have...as 70% sounds quite a bit off of what I am hearing...but whatever.

I highly doubt that zen would use a standard Glofo 14LPP process. they have always used specialized processes in the past as you know the 28nm process amd uses for Carrizo is different than the process used by early 28nm chips labled as low power /mobile application. expect something like 14HPP as LPP is designed for sub 35W chips and I expect a roughly 100W tdp for 14nm parts, possibly 150W for the 16 thread monster.

I always assumed that Glofo would offer a tweaked node to AMD. But when I asked him about this issue about one month ago, his reply was:

There has been rumors that GlobalFoundries would be making a "tuned" version of the LPP, but I haven´t heard anything like that from anybody. A tuned version doesn´t bode well either with the fact that AMD has stated that they have abandoned "custom nodes".

 

[juanrga]

Should we already have an Exact release date by now? something tells me it may not even be available till 2017.

I don`t care the clock speed, as long as it at least competes with Haswell... for all i care it could be clocked to 2GHZ max.

AMD is expected to give more details next year, including release date.

 

[etayorius]

Should we already have an Exact release date by now? something tells me it may not even be available till 2017.

I don`t care the clock speed, as long as it at least competes with Haswell... for all i care it could be clocked to 2GHZ max.

AMD is expected to give more details next year, including release date.

Ohh cool, thanks juanrga.

 

[juanrga]

The SPARC M7 ships at 4.1Ghz with 32 cores. There is little reason the 14/16nm parts which use a similar 20nm BEOL can't hit similar targets with just 8 cores.

TSMC 20nm high performance node != Samsung/Glofo 14LPE/LPP

You do realize that the process AMD is using is a tuned version of the samsung node...right?

This is not just off the shelf 14nm LPP. I also question what "insider foundry knowledge" others would claim to have...as 70% sounds quite a bit off of what I am hearing...but whatever.

Another poster with inside foundry knowledge also claims that AMD is using LPE/LPP

It is same as Samsung. Maybe some minor tweaking to take into account tool setups, but otherwise identical.

He also has a good record.

 

[salgado18]

AMD Zen FX CPU tops out at 8-core; Broadwell-E CPU tops out at 10-core. The only place where AMD seems to push a moar cores strategy is on servers: Zen Opteron up to 32-core; Broadwell Xeon up to 24-core.

AMD Zen FX CPU is rated up to 95W. Haswell-E is up to 140W. Broadwell-E seems to maintain same wattage.

I highly doubt it. I did find the news and rumours that said so, but none of them seemed to point to a reliable source.
And if you think about it, 8-core Haswell-E reaches 140W on 22nm, what's the chance that Intel, with all their research towards efficiency, will manage to fit an 8-core 14nm CPU in a 95W envelope? What about the 10-core Broadwell-E? That should be 140W as well, since it seems the market is ok with that number.

Based on that, I believe the 95W rumour is for a 4-core Zen. AMD is not yet an expert in the new arch, neither in the new process, to match Intel's power consumption. The 8-core Zen will probably reach 140W as well, or 125W if they do a good job and/or use lower clocks.

 

[jimmysmitty]

AMD Zen FX CPU tops out at 8-core; Broadwell-E CPU tops out at 10-core. The only place where AMD seems to push a moar cores strategy is on servers: Zen Opteron up to 32-core; Broadwell Xeon up to 24-core.

AMD Zen FX CPU is rated up to 95W. Haswell-E is up to 140W. Broadwell-E seems to maintain same wattage.

I highly doubt it. I did find the news and rumours that said so, but none of them seemed to point to a reliable source.
And if you think about it, 8-core Haswell-E reaches 140W on 22nm, what's the chance that Intel, with all their research towards efficiency, will manage to fit an 8-core 14nm CPU in a 95W envelope? What about the 10-core Broadwell-E? That should be 140W as well, since it seems the market is ok with that number.

Based on that, I believe the 95W rumour is for a 4-core Zen. AMD is not yet an expert in the new arch, neither in the new process, to match Intel's power consumption. The 8-core Zen will probably reach 140W as well, or 125W if they do a good job and/or use lower clocks.

Intel has also been known to highly over estimate their TDP and uses TDP vs AMD who has been known to either hit their TDP or use ADP instead of TDP so a 95W TDP (might be ADP) 8 core on 14nm might be possible.

 

[juanrga]

AMD Zen FX CPU tops out at 8-core; Broadwell-E CPU tops out at 10-core. The only place where AMD seems to push a moar cores strategy is on servers: Zen Opteron up to 32-core; Broadwell Xeon up to 24-core.

AMD Zen FX CPU is rated up to 95W. Haswell-E is up to 140W. Broadwell-E seems to maintain same wattage.

I highly doubt it. I did find the news and rumours that said so, but none of them seemed to point to a reliable source.
And if you think about it, 8-core Haswell-E reaches 140W on 22nm, what's the chance that Intel, with all their research towards efficiency, will manage to fit an 8-core 14nm CPU in a 95W envelope? What about the 10-core Broadwell-E? That should be 140W as well, since it seems the market is ok with that number.

Based on that, I believe the 95W rumour is for a 4-core Zen. AMD is not yet an expert in the new arch, neither in the new process, to match Intel's power consumption. The 8-core Zen will probably reach 140W as well, or 125W if they do a good job and/or use lower clocks.

If 10-core BDW is ~140W, then 8-core BDW would be ~ 112W. Zen likely has less L3 cache (20MB vs 16MB) which reduces power consumption. Reduce clocks by ~10% and we are on 95W easily.

95W for a quad-core Zen on 14nm would be a complete disappointment.

 

[juanrga]

https://www.semiwiki.com/forum/content/5180-globalfoundries-visit-%C2%96-part-2-%C2%96-waking-sleeping-giant.html

As interesting as all of that is, they then went on to describe their new FX-14 ASIC offering utilizing the 14LLP process manufactured at Fab 8 in Malta. As an outside observer one concern I had for GF was how they would differentiate themselves from Samsung and become more than a second source when they are licensing the 14nm process from Samsung. While we were in Burlington we heard the AMD has committed to 14LPP giving GF a volume first source customer to drive yield learning.

 

[Vogner16]

If 10-core BDW is ~140W, then 8-core BDW would be ~ 112W. Zen likely has less L3 cache (20MB vs 16MB) which reduces power consumption. Reduce clocks by ~10% and we are on 95W easily.

95W for a quad-core Zen on 14nm would be a complete disappointment.

Maybe to you I plan to overclock the living F#$% out of mine id be happy to hit 200W when its all said and done.

 

[jimmysmitty]

If 10-core BDW is ~140W, then 8-core BDW would be ~ 112W. Zen likely has less L3 cache (20MB vs 16MB) which reduces power consumption. Reduce clocks by ~10% and we are on 95W easily.

95W for a quad-core Zen on 14nm would be a complete disappointment.

Maybe to you I plan to overclock the living F#$% out of mine id be happy to hit 200W when its all said and done.

Overclocking does also depend on a lot of factors. the uArch is one but there are other factors such as the process node and how well it handled the extra voltage, how much extra voltage it can take and the heat dissipation.

Intels 65nm was better than AMDs 65nm at overclocking.

I guess we just need to wait and see what they finally come out as the final process node and how well it handles higher clocks.

 

[Vogner16]

If 10-core BDW is ~140W, then 8-core BDW would be ~ 112W. Zen likely has less L3 cache (20MB vs 16MB) which reduces power consumption. Reduce clocks by ~10% and we are on 95W easily.

95W for a quad-core Zen on 14nm would be a complete disappointment.

Maybe to you I plan to overclock the living F#$% out of mine id be happy to hit 200W when its all said and done.

Overclocking does also depend on a lot of factors. the uArch is one but there are other factors such as the process node and how well it handled the extra voltage, how much extra voltage it can take and the heat dissipation.

Intels 65nm was better than AMDs 65nm at overclocking.

I guess we just need to wait and see what they finally come out as the final process node and how well it handles higher clocks.

True, point was I don't care about how power efficient their top end 8 core cpu is, as I don't pay the power bill. If zen ends up with 150W tdp and overclocks to use 200W I wont complain.

I think that's what intel was thinking from their top end parts as well. most people with 6700K's overclock them to use 140W and if you noticed their stock tdp has increased in the last generation. desktops generally don't need efficiency as most people with i7 6700K's are running 600W plus power supplies and can easily run the power.

even amd's last gen 8350 took over 8 years for the average consumer to pay off the price difference in power draw compared to owners of the 3770K.

https://www.youtube.com/watch?v=fBeeGHozSY0


EDIT: point is power efficiency is irrelivent in the desktop market for 90% of gamers and power users.

U series and laptops... different story, desktop no.